Token-light
Retrieve less, but retrieve better.
TinySuite favors small ranked evidence sets over raw dumps. Less noise means more room for instructions, user context, and actual reasoning.
Philosophy
Smaller context. More useful agents.
TinySuite is built by Marc around a simple idea: AI tools should retrieve carefully, stay inspectable, and help local agents work without pretending context is free.
Core belief
The best context is usually smaller than the easiest context.
TinySuite tools search, filter, chunk, rerank, and return only the evidence an agent can actually use.
Why it exists
Most agent tools are built like context is free.
Context is not just a technical limit or a billing detail. It shapes how fast an agent responds, how much a local model can reason over, and whether the final answer can be tied back to real sources.
TinySuite exists for developers who want practical local agents without dumping whole webpages, transcripts, or histories into every prompt. It treats context as a budget and retrieval as a responsibility.
What makes it different
Small tools instead of a heavy platform.
TinySuite is not trying to become a hosted dashboard, account system, or all-in-one AI workspace. The tools are meant to sit close to your local agent workflow and do one job clearly.
Local-first
Keep the workflow under user control.
TinySearch runs as a self-hosted MCP research layer. It does not require a hosted dashboard or account just to give an agent useful web context.
Source-grounded
Answers should point back to evidence.
The model receives compact snippets with source URLs attached, so the final answer can cite what mattered instead of improvising from a messy page.
Composable
Search is a dependency, not the product.
TinySuite uses simple MCP shapes so agents can search, inspect known URLs, get current time, and keep the client routing easy to understand.
How it helps
Make local agents more capable without making them heavier.
For builders
TinySuite gives developers a narrow research loop they can inspect: search, crawl, chunk, rerank, and hand the model a grounded prompt.
For users
Smaller prompts can be faster, easier to verify, and less likely to bury the answer in irrelevant text.
For local models
Local LLMs get more useful when their limited context windows are filled with relevant evidence instead of boilerplate, navigation, and repeated page noise.
Ethics
Ethical AI tooling starts with restraint.
TinySuite does not promise magic or hide the retrieval process behind a black box. Its ethical posture is practical: keep tools small, preserve sources, avoid unnecessary hosted surfaces, and make the selected context possible to inspect.
User control
Prefer local, configurable infrastructure over workflows that require an account first.
Evidence over vibes
Return source-backed snippets so answers can be checked against the material used.
Less waste
Use fewer tokens by selecting relevant context instead of flooding prompts with noise.
Marc's direction
TinySuite grows by staying small.
TinySearch is the first launched product. TinyContext is planned as a local-first memory layer that follows the same philosophy: save useful facts, retrieve selectively, and stay honest about what memory can and cannot do.
The point is not to collect features. The point is to make local agent workflows easier to trust, easier to inspect, and easier to run.